Biologically-inspired memory for AI agents.

Cortex combines principles from neuroscience, immunology, and library science into a unified memory system that actually works across sessions.

Every AI memory system picks one insight and stops. Mem0 does extraction. Zep does temporal graphs. Letta does self-management. Cortex combines all of them — because that's how actual memory works.

• Three memory types: Episodic (events, decays fast), Semantic (facts, durable), Procedural (learned behaviors, most durable)
• Consolidation "sleep cycles": Raw episodes get compressed into semantic facts. Each re-encounter makes facts more precise, not just more confident.
• Immune system decay: Memories require retrieval to survive. No retrieval = decay. Contradictions get detected and pruned.
• Multi-strategy retrieval: Semantic similarity + BM25 keyword + graph traversal + temporal reasoning, fused with Reciprocal Rank Fusion.
• "Walk into the room": One call loads a compressed, task-aware context bundle — call prep surfaces risks, bug triage surfaces product issues.
• Fully local: SQLite + local embeddings. Zero API calls. Your data never leaves your machine.
• MCP-native: Works with Claude Code, Cursor, or any MCP client.

pip install cortex-memory

# Store memories
cortex remember "Acme Corp renewed for $120k ARR" -t "customer:acme"
cortex remember "John wants API access by Q3" -t "customer:acme" -s "meeting"

# Search
cortex search "Acme contract"

# Knowledge graph
cortex graph "Acme Corp"

# Health dashboard
cortex status

# Start MCP server (for Claude Code / Cursor)
cortex serve

Add to your Claude Code config:

{
  "mcpServers": {
    "cortex": {
      "command": "cortex",
      "args": ["serve"]
    }
  }
}

Then use cortex_remember, cortex_recall, cortex_forget, cortex_connect, cortex_traverse, cortex_entities, cortex_health, cortex_pin tools from any agent.

Input → Encode → Episodic Store → [Consolidation] → Semantic Store → [Decay] → Archive
                                                   → [Contradiction] → Supersede
                                → [Pattern Extract] → Procedural Store → [Track] → Refine

Storage: Single SQLite file at ~/.cortex/cortex.db with sqlite-vec for vector search.

Embeddings: Local via sentence-transformers (default: all-MiniLM-L6-v2). No API calls.

Decay model: Modified Ebbinghaus with retrieval-frequency boosting. Episodic: 3-day half-life. Semantic: 30-day. Procedural: 60-day. Pinned memories never decay.

MIT

Deway — the autonomous adoption platform for B2B SaaS. We run 35 AI agents daily and needed memory that actually works.